Module 09 Activity

Scenario

Spend an afternoon attacking your own application before someone else does.

What you build

A red-team exercise covering injection, exfiltration and non-adversarial misuse, reported honestly.

Steps

1. Run all seven probes by hand, including the long-prefix one. Record what each produced.
2. Plant an injected instruction in a document your indexer will pick up and see whether it changes an answer.
3. Work through the six exfiltration paths and check each against your own system - especially the cache, the trace store and any model-composed URLs.
4. List five ways a reasonable person could misuse the system, and check whether the interface states its scope.
5. Make out-of-scope refusals redirect rather than merely decline.
6. Write the report with four sections: confirmed, not reproduced, not tested, accepted risk.

Evidence to hand in

• Seven probe results.
• The injection test and its outcome.
• The six exfiltration paths checked against your system.
• Five misuse cases and what the interface says about scope.
• The four-section report.

Review checklist

• The long-prefix probe was actually run.
• The injection test used a document the indexer picks up, not a pasted prompt.
• The cache, trace store and URL paths were all checked.
• Out-of-scope refusals redirect somewhere specific.
• not_tested and accepted_risk are both populated.

Mark complete